Why did we create an AI that can negotiate?
It was a beautiful day in France. Springtime was well on the way.
I could not marvel at it though. I was in a meeting room negotiating with seasoned French businessmen. Why are they so different from the Germans, I thought? And why are the Germans so different from Ukrainians and Spaniards? The French executives deemed my clever proposals grossly inadequate and closed the door on my way out. Does this have to be so painful? After two years of this, I thought there is no cognitive bias or cultural difference I have not painfully experienced. Could technology help here? I was not sure.
Fast forward 1.5 years. Friends and former colleagues Kaspar Korjus and Kristjan Korjus called me to ask advice on a potential startup. My sole mission became to persuade Kaspar and Kristjan to stop and to start something in negotiations instead.
Kaspar was just a role model in solving something like this. He was the MD of e-Residency, one of the most ambitious government tech projects in Europe. He had persuaded the police, the Foreign Office, the prime minister, the president and all ministries to invest government money to open Estonia’s sensitive electronic infrastructure to all foreigners in the world so they could start companies in the European Union without ever having to visit. Crazy, right? That’s some negotiation skills.
Kristjan, Kaspar’s brother, is a PhD mathematician who was building artificial brains for autonomous delivery robots at Starship. His PhD thesis was about deep learning agents aligning their values and collaborating. He led a team that replicated DeepMind’s original work that made waves in scientific media and his team members went to work for Elon Musk’s OpenAI and to DeepMind.
With Kaspar and Kristjan we had the strongest team for automating negotiation in the world. So we set about raising money. Our first investor was Jaan Tallinn — AI philanthropist and founder of Skype and investor in DeepMind. He was the most intelligent person we had the privilege to speak to, ever. He could finish our sentences on reaching Pareto efficient outcomes through value alignment. This was powerful. Then came Taavet Hinrikus (Founder and CEO of Transferwise), Ott Kaukver (CTO of Twilio) and Sten Tamkivi (ex MD of Skype and a CPO of Topia) followed. We can proudly say we have one of the best team of angel investors one could find when fundraising in Europe.
We knew what we had to do. We wanted to create a system that could negotiate without human biases on a superhuman level. It had to be able to learn from each negotiation and get better in closing Pareto efficient deals. These are agreements where one side cannot get a better deal without hurting the other party.
Initially, we wanted to create a global sparring platform for negotiators and teach the AI based on these learnings. We took Harvard Business School negotiation simulations and asked people to negotiate. What we quickly noticed was that no matter how proficient negotiators they were, they almost never achieved Pareto optimum. There was an effect that made the situation even worse. According to the Pareto principle, 80% of commercial agreements are ‘long tail’- meaning low value, high volume. Negotiating this vast number of contracts in large corporations is almost impossible with people. Enterprises even have a name for them — the unmanaged long tail.
Achieving superhuman levels in a situation where the largest companies in the world called 80% of their contracts ‘unmanaged’ became easier than we thought. We don’t need to build a bot with general intelligence. We just need to build a system that is better than the status quo- which is people avoiding having to negotiate low-value high volume deals at any cost.
So we built a bot that can negotiate. Our system consists of two parts: the value function and the negotiation flow. The value function defines every possible variation and all dependencies between the negotiation items. The negotiation flow is a collection of recursive graphs of human-understandable negotiation strategies and tactics. The role of AI is to learn from each negotiation and get better in the next. Mathematically the shortest way to a Pareto optimal deal is by creating value for both sides. Note that AI will learn about humans not from humans during these negotiations. The negotiation strategies and tactics are prepared by our negotiation scientists. Thus it does not perpetuate existing human biases. The task of AI is to find the best way to get to a Pareto efficient deal.
Pactum’s first negotiation was for an online limousines services aggregator. The bot negotiated with an elderly limousines rental company owner in Spain. Pactum was able to complete the negotiation in 8 minutes. When we asked for feedback the gentlemen said that he hates computers, but this system was really nice and polite to him. This gave us a lot of vigor to continue. We realized that negotiation is a source of stress for most people so they try to avoid it. They would more likely negotiate with a robot which is kind, polite, professional, does not push the other party into concessions and does not have an ego to deal with.
We’re now focusing on helping Fortune 500 type companies pick up locked value from their long tail agreements while increasing value for the other side. People can thus focus on important deals where human ingenuity is always needed.
